Asset Pricing Implications of Learning about Long-Run Risk

Asset Pricing Implications of Learning about Long-Run Risk Daniel Andrei Michael Hasler Alexandre Jeanneret August 15, 016 Abstract We develop a dynamic pure-exchange economy that nests learning about the level of expected economic growth and learning about its persistence. Using GDP growth and forecast data, we find that a model of learning about persistence is more likely to prevail in the U.S. economy than a model of learning about level. In equilibrium, learning about persistence generates strong fluctuations in stock-market volatility, risk premium, and Sharpe ratio. When persistence is uncertain, these moments become counter-cyclical and the effect of long-run risk is amplified. We find empirical support for these theoretical predictions and provide evidence that learning about persistence helps explain the conditional properties of asset returns. We are grateful to Michael Brennan for helpful comments, which have led to substantial improvements in the paper. We also would like to thank Pat Akey, Mariana Khapko, and Chayawat Ornthanalai for their suggestions. We thank UCLA, the University of Toronto, and HEC Montréal for financial support. UCLA Anderson Graduate School of Management, 110 Westwood Plaza, Suite C40, Los Angeles, CA 90095, USA; +1 (310) 85-3544; daniel.andrei@anderson.ucla.edu; www.danielandrei.info. University of Toronto, Rotman School of Management, 105 St. George Street, Suite 431, Toronto, ON M5S 3E6, Canada, michael.hasler@rotman.utoronto.ca; www.rotman.utoronto.ca/faculty/hasler. Department of Finance, HEC Montréal, 3000 Côte-Sainte-Catherine, Montréal, QC H3T A7, Canada; +1 (514) 589-6699; alexandre.jeanneret@hec.ca; www.alexandrejeanneret.com. 1

1 Introduction The long-run risk model of Bansal and Yaron (004) has gained increasing popularity as a potential explanation for a broad class of asset market phenomena. Its two building blocks are recursive preferences (Epstein and Zin, 1989; Weil, 1989) and the existence of a small, persistent component in expected consumption growth. One potential caveat is the assumption that investors have perfect knowledge about the existence and behavior of this small component. Arguably, neither the degree of persistence nor the level of expected consumption growth is directly observable to investors. It is important to understand whether relaxing the perfect knowledge assumption would qualitatively change the asset pricing implications of the traditional long-run risk model. This is the objective of this paper. We propose a general equilibrium model with incomplete information in which a representative agent learns about key parameters of the consumption growth process. In particular, two dimensions are simultaneously unobservable to the agent: the level of the expected consumption growth and its persistence. While uncertainty about the level of expected consumption growth has been thoroughly analyzed in the literature, 1 we show that learning about the second dimension that is, learning about long-run risk generates distinct, important implications for the dynamics of asset returns. In particular, learning about long-run risk implies significant time-series variation in stock-market volatility, risk premium, risk-free rate, and Sharpe ratio. Moreover, when uncertainty about the degree of persistence in consumption growth is present, all these market outcomes become tightly linked to current economic conditions, thereby rationalizing why stock market volatility, risk premia and the Sharpe ratio tend to increase as economic conditions deteriorate. Learning about the unobservable degree of persistence in consumption growth can improve our understanding of asset pricing facts that have remained difficult to explain with traditional asset pricing models. To be more precise, our model assumes away any variation in the volatility of consumption growth an important source of risk causing a time-varying equity premium in the long-run risk model (Bansal, Kiku, and Yaron, 01; Beeler and Campbell, 01); or time-varying risk of disasters (Gabaix, 01; Wachter, 013); or any utility specification that induces time-varying risk aversion (Campbell and Cochrane, 1999). Instead, in our model, all business-cycle fluctuations in asset pricing moments arise endogenously due to investors updating about the unobservable degree of long-run risk. We start by deriving a general specification which nests a continuum of learning choices. A unique learning parameter, which takes values between zero and one, specifies the type of learning that takes place in the market. At one extreme, the only unobservable dimension is the 1 See David (1997), Veronesi (1999, 000), Brennan and Xia (001), David and Veronesi (00) among many others, and excellent surveys by Pastor and Veronesi (009) and Ziegler (01).

level of expected consumption growth this corresponds to the case of learning about economic growth. At the other extreme, the only unobservable dimension is the degree of persistence in consumption growth this corresponds to the case of learning about long-run risk. In between, both types of learning are simultaneously present in the economy and the learning parameter controls their relative importance for investors. We estimate this parameter to understand which type of learning best explains the dynamics of economic fundamentals in the United States. A Kalman-filter Maximum Likelihood estimation (Hamilton, 1994) over the period Q4:1968 to Q4:015 on real GDP growth and analyst forecast data shows that learning is exclusively about long-run risk (and not about the level of expected economic growth); statistically, we cannot reject the null that learning is only about long-run risk. The estimation indicates that agents perceived degree of long-run risk (i.e., their estimate of the persistence parameter) varies significantly over time, in contrast with the traditional assumption of long-run risk models that the degree of long-run risk remains constant across time and economic conditions. Based on the results of the estimation, we focus our analysis on the special case of the model in which agents learn exclusively about long-run risk. The main prediction is that uncertainty about long-run risk greatly enhances the sensitivity of stock market volatility, the equity risk premium, and the Sharpe ratio to current economic conditions and to the perceived degree of long-run risk. The mechanism through which this occurs relies on how agents form expectations and react to news about expected economic growth. To grasp the intuition, let us assume that agents receive news about expected economic growth from the Survey of Professional Forecasters, as publicly available from the Federal Reserve Bank of Philadelphia. Without loss of generality, consider that the latest consensus forecast on future economic growth is lower than previously thought, which we interpret as bad news. During a recession, such negative news prompts agents to expect a longer downturn and, therefore, to perceive more long-run risk. Stock prices then drop because of both bad news and the increased long-run risk. In contrast, during an economic boom, the same piece of news induces agents to perceive less long-run risk, through an identical Bayesian updating mechanism, which dampens the initial effect of the unexpected deterioration in analyst forecasts. Stock returns therefore respond differently to news about future economic growth in bad versus good economic times. This learning mechanism generates significant time-series variation in all asset pricing quantities. Notably, this effect is amplified when uncertainty about long-run risk increases, yet it disappears when agents have perfect information about the degree of persistence. Our theoretical model predicts that uncertainty about long-run risk is a key channel There are other (much weaker) effects that take place. We fully describe and analyze them in the paper. 3

through which asset pricing outcomes vary with both the state of the economy and the perceived degree of long-run risk. We test this prediction with S&P 500 data and find that all the regression coefficients predicted by the model have the correct sign and are strongly significant. Furthermore, the model-implied price-dividend ratio, stock return volatility, risk premium, and Sharpe ratio explain a large fraction of the variation in their observed counterparts with high coefficients of determination. Our analysis thus provides strong empirical support for a model with learning about long-run risk. In comparison, the asset pricing implications differ substantially when we consider an alternative version of the model with complete information about the consumption growth process or with incomplete information about the expected consumption growth only. In particular, these specifications are unable to generate time-variation in volatility, risk premium, and Sharpe ratio, at odds with the data. Beyond the literature that contributes to our understanding of stylized asset-pricing facts in frictionless settings with rational expectations (Campbell and Cochrane, 1999; Bansal and Yaron, 004; Gabaix, 01; Wachter, 013), this paper also relates to the literature on learning in financial markets. 3 Most studies in this literature commonly assume that parameters are known and that the unobservable dimension is the level of expected consumption growth, with a few notable exceptions. Johannes, Lochstoer, and Mou (016) consider different Markov switching models with both unknown parameters and unknown states and show that model and state uncertainty help rationalize the observed levels of return volatility, risk premium, and Sharpe ratio. 4 They build a theoretical model in which the agent has anticipated utility (Kreps, 1998; Cogley and Sargent, 008), and thus parameter uncertainty is not a priced risk factor. In contrast, we focus on the dynamics of asset returns in a fully rational setup with Epstein-Zin utility and show, both theoretically and empirically, that uncertainty about the degree of persistence in consumption is a key channel through which asset pricing moments depend on economic conditions. Collin-Dufresne, Johannes, and Lochstoer (016) show that parameter uncertainty generates endogenous long-run risk and therefore implies a large equilibrium risk premium when the representative agent has a preference for early resolution of 3 Although a comprehensive review of this literature is beyond the scope of our paper, there are a few closely related studies. Veronesi (1999) shows that return volatility and the risk premium are hump-shaped functions of the state of the economy when the representative agent learns about a discrete-state expected consumption growth rate. Ai (010) considers a production economy with learning about productivity shocks and Epstein-Zin preferences. He shows that the relation between information quality and the risk premium is negative, whereas it is positive in pure-exchange economies (Veronesi, 000). Croce, Lettau, and Ludvigson (015) show that, when the representative agent has a preference for early resolution of uncertainty, learning about expected consumption growth in a bounded rationality limited information model generates both a large risk-premium and a downward-sloping term structure of risk. 4 Pakoš (013) analyzes a similar economy featuring a three-state Markov chain in which a representative agent cannot distinguish between a mild recession and a lost decade. This modeling choice automatically introduces a stronger response to news in bad times. In our case, the agent learns about long-run risk at all times good or bad and the asymmetric stock price sensitivity across economic conditions arises endogenously from the learning mechanism. See also Johnson (001) for a model with learning about persistence. 4

uncertainty. One key result in Collin-Dufresne et al. (016) is that the impact of parameter uncertainty on asset prices varies substantially, depending on which parameters are considered. Therefore, it is important to understand which parameters when unknown matter most to investors, a question that we address in two ways. First, we provide new empirical evidence that agents are significantly more likely to learn about the degree of persistence in expected consumption growth. Second, our theory offers insights into how this learning choice generates novel interactions between the level of long-run risk, its uncertainty, and the state of the economy. Overall, these findings improve our understanding of the time variation and counter-cyclicality of return volatility, risk premia, and Sharpe ratio observed in the data. The remainder of the paper is organized as follows. Section introduces a model that nests learning about economic growth and learning about persistence. Section 3 calibrates the model and provides empirical evidence that investors essentially learn about the persistence of expected consumption growth. Section 4 presents our theoretical predictions, while Section 5 tests these predictions. Section 6 offers some concluding remarks and future directions for research. Model This section analyzes a pure-exchange economy in which the representative agent faces incomplete information about the dynamics of consumption. Bayesian learning allows the agent to reduce the uncertainty about the level of economic growth, the severity of long-run risk, or a combination of the two. We derive the equilibrium asset pricing implications based on these different learning choices..1 Environment Consider a pure-exchange economy defined over a continuous-time horizon [0, ) and populated by a representative agent who derives utility from consumption. The agent has stochastic differential utility (Epstein and Zin, 1989) with subjective time preference rate β, relative risk aversion γ, elasticity of intertemporal substitution ψ, and an indirect utility function given by J t = E t [ where the aggregator h is defined as in Duffie and Epstein (199): h(c, J) = ( β 1 1/ψ t ] h(c s, J s )ds, (1) C 1 1/ψ (1 γ)j [(1 γ)j] 1/φ 1 ), () 5

with φ 1 γ. Note that if γ = 1/ψ, then φ = 1 and we obtain the standard CRRA utility. 1 1/ψ A risk-free asset is available in zero net supply. A single risky asset the stock is available in unit supply, and represents the claim to the aggregate consumption stream, which follows the process dδ t δ t = µ t dt + σ δ dw δ t, (3) where dw δ is a standard Brownian motion. The expected consumption growth rate, µ t, is unobservable. The history of the consumption process (3) provides an incomplete signal about the expected growth rate. Additionally, we assume that the agent continuously receives a time-varying forecast of µ t, produced form a Survey of Professional Forecasters. The expected growth rate then becomes µ t = f t + b, (4) where f t is the observed forecast and b is an unknown bias, which we assume for simplicity to be constant. 5 The expected growth forecast follows where the Brownians dw f and dw δ are independent. 6 df t = Λ( f f t )dt + σ f dw f t, (5) The mean-reverting parameter Λ, which controls the persistence of the forecast process and thereby of the expected growth rate is assumed to be unobservable. We assume that the agent starts with a prior λ on Λ. Our aim is to embed the above two dimensions of uncertainty about the level of the expected growth rate, µ t, but also about its degree of persistence, Λ in a unified theoretical framework. To this end, we propose the following specification: dδ t δ t = [f t + (1 θ)l]dt + σ δ dw δ t (6) df t = (θλ + λ)( f f t )dt + σ f dw f t, (7) where (1 θ)l and θλ + λ are respectively the bias b and the parameter Λ from (4) and (5). 5 In Appendix C.1, we allow this bias to be time-varying and driven by a standard Brownian motion. We show that this richer specification does not change our theoretical results and it has only a minor impact on the calibration of the model, which now features an additional parameter to be estimated. 6 The model can be extended by allowing the two Brownians in (3) and (5) to be correlated. However, we assume that they are independent. This simplifies the description of the model without changing our main message. 6

The parameter θ, which is novel to our formulation, allows us to contrast the equilibrium implications of two types of learning. When θ is either 0 or 1, only one dimension of uncertainty is present and the agent focuses her attention on one type of learning only. If θ = 0, the agent has perfect knowledge about the magnitude of long-run risk, which is now determined by λ. However, she now faces uncertainty about the level of the expected economic growth rate, which in this case becomes the growth forecast f t plus an unknown bias (1 θ)l. That is, the agent views the growth forecast as imperfect and seeks to improve this estimate; we call this case learning about economic growth. Conversely, if θ = 1, the agent believes that the forecast f t is a perfect estimator of expected consumption growth. However, she now faces uncertainty about the mean-reverting speed of the consumption growth rate, λ+λ, and consequently aims to estimate it; we call this case learning about long-run risk. Finally, the parameter θ can take any value between 0 and 1, in which case the agent learns simultaneously about economic growth and long-run risk. If θ approaches zero, the agent is more concerned with learning about the level of economic growth; if θ approaches one, the agent is more concerned with learning about the risk that she faces in the long term due to the slow-moving persistent component. The specification (6)-(7), therefore, conveniently embeds a continuum of learning choices in one unified framework, while preserving the linearity of the learning exercise. 7. Learning dynamics We now examine the agent s updating of beliefs as new information arrives. The agent starts with the following priors about the unknown parameters l and λ: [ ] ([ ] [ ]) l 0 ν l,0 0 N,. (8) λ 0 0 ν λ,0 With this specification, the value of θ affects neither the agent s prior about the long-term level of the expected growth rate, which always equals f, nor the agent s prior about the mean-reverting speed of the expected growth rate, which always equals λ. Notice that there is no correlation between the two priors. Define F t the information set of the agent at time t, and denote by l t E[l F t ] the estimated parameter l and its posterior variance by ν l,t E[(l l t ) F t ]. Similarly, denote by λ t E[λ F t ] the estimated parameter λ and its posterior variance by ν λ,t E[(λ λ t ) F t ]. 7 The ingredient of the model which helps preserve the linearity of the learning exercise is the absence of the bias from the process (7). In other words, we assume that the bias in Professional Forecasters estimate of the expected growth rate is not following the same process as the forecast itself. Instead, in the specification (6), the bias (1 θ)l is constant. We also allow for i.i.d. changes in l in Appendix C.1, where we show that the structure of our model is preserved, while the calibration does not change significantly. 7

These estimates and posterior variances are such that l N( l t, ν l,t ), λ N( λ t, ν λ,t ), (9) where N(m, v) denotes the Normal distribution with mean m and variance v. It is convenient to interpret the estimates l and λ as the filters and the two posterior variances ν l,t and ν λ,t as the levels of uncertainty. Following Liptser and Shiryayev (1977), the filters evolve according to (see Appendix A): [ ] d l t = d λ t [ (1 θ)νl,t σ δ 0 0 θ( f f t)ν λ,t σ f ] [ ] dŵ t δ dŵ f, (10) t where dŵ δ t = 1 σ δ ( dδt δ t ) [f t + (1 θ) l t ]dt, dŵ f t = 1 (df t (θ λ t + σ λ)( f ) f t )dt, (11) f are independent Brownian motions, resulting from the filtration of the agent. For clarity, we will hereafter use the term consumption growth shocks to refer to dŵ δ innovations and the term expected consumption growth shocks to refer to dŵ f innovations. One can observe from (10) that the agent does not update λ if θ = 0 and thus keeps her prior λ regarding the degree of long-run risk in the economy. Conversely, if θ = 1, the agent does not learn about l and thus uses the growth forecast f t as the predictor of future consumption growth. Finally, a value of θ between zero and one implies that the agent updates both estimates l and λ. The estimate related to expected economic growth, l, is perfectly positively correlated with consumption δ, which implies extrapolative expectations (Brennan, 1998): after positive (negative) consumption shocks, the agent revises her estimate of the expected growth rate upwards (downwards). More interestingly, this extrapolative expectation formation also occurs in the case of learning about long-run risk. However, the effect now depends on the state of the economy, which is characterized by the level of the current growth rate forecast f t relative to its unconditional mean f. In good times ( f f t < 0), positive expected consumption growth shocks decrease the agent s estimate of λ; in bad times ( f f t > 0), negative expected consumption growth shocks decrease the agent s estimate of λ. In both situations (i.e., positive shocks in good times or negative shocks in bad times), the agent extrapolates that expected consumption growth becomes more persistent, and that long-run risk increases. As we will show, this extrapolative expectation formation plays a critical role in the behavior of equilibrium asset prices. 8

The dynamics of the posterior uncertainties about l and λ are respectively given by dν l,t = (1 θ) νl,t dt, σδ dν λ,t = θ ( f f t ) νλ,t dt, (1) σf and indicate that both uncertainties converge to zero in the long-run, since the agent learns about constants. 8 Notably, the convergence is faster for ν λ when f t is far away from f. This is because news about λ become more informative when the economy is either in very good or very bad times..3 Asset pricing We turn now to equilibrium asset prices and examine the implications of learning about economic growth and/or about long-run risk..3.1 Equilibrium In this environment, the equilibrium is standard and the technical details are relegated to Appendix B. Solving for the equilibrium involves writing the HJB equation for problem (1): max {h(c, J) + LJ} = 0, (13) C with the differential operator LJ following from Itô s lemma. We guess the following value function (Benzoni, Collin-Dufresne, and Goldstein, 011): J(C, f, l, λ, ν l, ν λ ) = C1 γ 1 γ [βi(x)]φ, (14) where I(x) is the price-dividend ratio, and x [f l λ ν l ν λ ] denotes the vector of state variables, whose dynamics are given in (7), (10), and (1), respectively. Substituting the guess (14) in the HJB equation (13) and imposing the market clearing condition, C = δ, yields the partial differential equation (45) for the log price-dividend ratio. We solve for this equation numerically using Chebyshev polynomials (Judd, 1998). The PDEs corresponding to the two polar forms of learning (i.e., θ = 0 and θ = 1) are provided respectively in Equations (46) and (47) in Appendix B. 8 It is straightforward to generate positive steady-state uncertainty by assuming that learning is regenerated, i.e., that l and λ are not constants but move over time (we do this for the parameter l in Appendix C.1). This would unnecessarily complicate our setup without affecting our qualitative implications. Moreover, as shown by Collin-Dufresne et al. (016), learning about a constant generates substantial effects when the representative agent has a preference for early resolution of uncertainty, as it is the case in our setup. 9

In order to characterize the effects of learning on equilibrium outcomes, we propose the following conjecture. Conjecture 1. With preferences satisfying γ > 1 > 1/ψ, we expect: I f I > 0, I l I > 0, I λ I > 0, I νl I < 0, I νλ I < 0. (15) This conjecture, which we will verify with a numerical exercise in Section 4, is likely to be true for a wide range of parameters. In fact, several inequalities follow directly from the definition of the value function in (14). Taking the derivative of J with respect to any of the five state variables yields J ( ) = φj I ( ) I, (16) with the product φj being positive when γ > 1 > 1/ψ. Then, note that due to non-satiation, expected lifetime utility must rise as investment opportunities improve, i.e., J f > 0 and J l > 0. Using (16), this implies the first two inequalities of Conjecture 1. Next, being risk averse, the agent prefers less uncertainty and thus J νl < 0 and J νλ < 0, which yields the last two inequalities of Conjecture 1. The only inequality that needs numerical validation is I λ/i > 0. Because the agent prefers early resolution of uncertainty, we expect that she prefers less persistence and thus J λ > 0. 9.3. State-price density, risk-free rate, and market price of risk Following Duffie and Epstein (199), the state-price density satisfies ξ t = e t 0 h J (C s,j s)ds h C (C t, J t ) = e t 0 [(φ 1)/I(xs) βφ]ds β φ C γ t I(x t ) φ 1, (17) and thus the dynamics of the state-price density follow dξ t ξ t = r t dt m t dŵt, (18) where r is the risk-free rate, m is the -dimensional market price of risk, and Ŵ [Ŵ δ, Ŵ f ] is the -dimensional standard Brownian motion. Letting σ I (x) [σ I1 (x) σ I (x)] be the diffusion vector of the price-dividend ratio and applying Ito s lemma to Equation (17) yields 9 In Appendix E, we perform a numerical evaluation of I λ/i for a wide range of values of the risk aversion, the intertemporal of elasticity substitution, and two state variables f t and λ t. We find that I λ/i is positive in all cases. We also elaborate on parametrizations for which this term can become negative, which could happen outside the standard calibration of our model. 10

the risk-free rate r t = β + 1 ψ [f t + (1 θ) l t ] γ+γψ ψ σ δ (1 φ) [ σ I1 (x t )σ δ + 1 (σ I1 (x t) + σ I (x t)) ] (19) and the market price of risk m t = [ γσ δ + (1 φ)σ I1 (x t ) (1 φ)σ I (x t )], (0) The two diffusion elements of the price-dividend ratio are given by σ I1 (x t ) = (1 θ)ν l,t σ δ I l I (1) I f σ I (x t ) = σ f I + θ( f f t )ν λ,t I λ σ f I. () Focusing on the equilibrium risk-free rate, the second term of the expression (19) indicates that fluctuations in expected consumption growth, which result from both the mean-reverting property of the growth forecast and from learning about the level of expected consumption growth, generate a procyclical risk-free rate. Furthermore, when the investor prefers early resolution of uncertainty (i.e., 1 φ > 0), the risk-free rate contains an additional term due to variations in f, l, and λ. This term arises because the agent dislikes long-run risk and the uncertainty regarding the level of such risk. The resulting effect is a lower risk-free rate due to greater demand for the safe asset. The market price of risk, defined in Equation (0), contains two elements, one for each of the two Brownians in the economy. The uncertainty about the expected growth rate, ν l,t, increases the first component when the agent learns about the level in expected growth (θ < 1). The impact of learning about long-run risk (θ > 0), instead, is present in the second component of the market price of risk and depends on the state of the economy, as given by the difference f f t. Following Conjecture 1, I λ/i > 0 and the market price of risk is expected to increase in bad times, when f f t > 0, and to decrease in good times, when f f t < 0. We will examine in depth this theoretical prediction in Section 4.3..3.3 Stock market volatility We now determine how the level and the dynamics of stock return volatility vary with different learning choices. The diffusion of stock returns, σ, satisfies σ t = [ ] σ δ + σ I1 (x t ) σ I (x t ), (3) 11

which, after replacing (1)-(), can be written as σ t = [ ] σ δ + (1 θ)ν l,t I l I σ δ σ f I f + θ( f f t)ν λ,t I λi I σ f. (4) The diffusion of stock returns indicates that learning about the level of expected consumption growth (θ < 1) can generate excess volatility in stock returns. According to Conjecture 1, the term I l/i is positive, and thus uncertainty about the expected growth rate ν l,t increases the magnitude of the first diffusion term. With uncertainty and learning about long-run risk, the sensitivity of stock returns to expected consumption growth shocks now varies with the current state of the economy, as determined by f f t. As we can see from the second diffusion term, the role of learning about long-run risk (θ > 0), due to the uncertainty associated with such risk (ν λ > 0), creates an asymmetric stock market response to shocks. To understand the effect, consider a negative shock on expected consumption growth, dŵ f t < 0, although the same intuition holds for a good shock. A bad shock in bad times induces the agent to update that there is more long-run risk, as she now believes that expected consumption growth becomes more persistent. The stock price thus drops not only because of the negative shock but also because of greater long-run risk. Hence, stock returns strongly react to shocks when the economy is in a bad state. By contrast, following a bad shock in good times, the agent perceives less long-run risk, which mitigates the initial effect of the negative shock. Learning about long-run risk thus attenuates the stock price response to shocks in good times, whereas it amplifies the response in bad times. Therefore, learning about the level of consumption growth helps generate a high, but constant, level of stock return volatility. In comparison, learning about long-run risk creates an asymmetric relation between stock returns and shocks that yields counter-cyclical stock return volatility. Our numerical analysis in Section 4. will illustrate and explore these theoretical predictions in greater detail..3.4 Equity risk premium We now examine how the equity risk premium varies with learning. The risk premium in our economy is defined as µ t r t = σ t m t. Using the expressions (0) and (4), we obtain µ t r t = ( σ δ + (1 θ)ν l,t σ δ ) ( I l γσ I δ + (1 φ) (1 θ)ν l,t σ δ ) ( I l I + (1 φ) σ f I f I + θ( f f t)ν λ,t σ f I λi ). (5) The equity risk premium consists of two terms. The first term is the risk premium arising from fluctuations in consumption growth. This part, which relates to the uncertainty about the level of consumption growth, becomes particularly relevant when the agent is strongly 1

risk-averse and prefers early resolution of uncertainty, i.e., when ψ > 1/γ. The second term captures the risk premium arising from fluctuations in expected consumption growth. When γ > 1/ψ, this term is positive and increases with the uncertainty about the persistence in expected consumption growth because the agent dislikes both long-run risk and the uncertainty about it. Furthermore, the term in brackets is higher (lower) when the growth forecast f t is below (above) its long-term level f. As a result, the equity risk premium in our economy fluctuates over time with learning about long-run risk. However, this risk premium remains independent from the state of the economy if the agent can perfectly estimate the persistence parameter, as ν λ,t becomes null. A similar prediction is obtained if the agent concentrates her attention to estimating the level of expected consumption growth, which means that θ = 0. Various types of learning, thus, will generate different asset pricing implications. We will thoroughly examine these implications in Section 4.. 3 Calibration We now calibrate the model and identify the type of learning that is more likely to explain the data. To this end, we use the mean analyst forecast on 1-quarter-ahead real GDP growth as a direct measure of the growth forecast f t and the realized real GDP growth as a proxy for the growth rate of the process δ t. We consider data on output rather than on consumption to exploit time series of analyst forecasts. Data are obtained from the Federal Reserve Bank of Philadelphia and are available at quarterly frequency from Q4:1968 to Q4:015. We use the dynamics of the filters l and λ (10), the dynamics of the uncertainties about l and λ (1), and the filtered Brownian shocks (11) to generate model-implied paths of consumption and the expected growth rate. We estimate the model by Kalman-filter Maximum Likelihood estimation (Hamilton, 1994) and determine the values of the parameters σ δ, f, σ f, λ, θ that provide the closest fit to realized observations. The initial priors are l 0 = λ 0 = 0, ν l,0 = 0.0, and ν λ,0 = 0.. The details specific to our implementation are summarized in Appendix C, while Table 1 reports the results. The estimation suggests that an environment with learning about the persistence of economic growth is most likely, as the learning parameter is θ = 0.99. While we clearly reject the null hypothesis that the learning parameter θ is equal to 0, we cannot reject the null that it is equal to 1 (p-value=0.99). Hence, the data indicate that learning is exclusively about the persistence (and not about about the level) of expected economic growth. We confirm this finding by comparing the results with two alternative specifications, which are presented in Table 1. First, we consider the specific case of θ = 1 and re-estimate this constrained version 13

Parameter Symbol Base case θ = 1 θ = 0 Output growth volatility σ δ 1.4% 1.4% 1.45% (5 10 4 ) (6 10 4 ) (6 10 4 ) Long-term expected growth f.59%.59%.61% (0.005) (0.005) (0.004) Expected growth volatility σ f.3%.3%.35% (7 10 4 ) (9 10 4 ) (9 10 4 ) Long-term mean reversion speed λ 0.78 0.78 0.95 (0.59) (0.55) (0.19) Learning parameter θ 0.99 (0.3) Akaike information criterion AIC, 484, 486, 480 Table 1: Parameter estimates This table reports the estimates of the model parameters. To calibrate the model, we use mean analyst forecast on the 1-quarter-ahead real GDP growth as a measure of the growth forecast and the realized real GDP growth as a proxy for consumption growth. The estimates are obtained by Maximum Likelihood for the period Q4:1968 to Q4:015, using data from the Federal Reserve Bank of Philadelphia. The table compares the estimation results of the base case model with those of two constrained models, with θ = 1 and θ = 0, respectively. Standard errors are in brackets and statistical significance at the 10%, 5%, and 1% levels is denoted by *, **, and ***, respectively. of the model. We obtain almost identical parameter values as in the base case. In addition, this estimation generates approximately the same goodness of fit (using the Akaike information criterion) as that of the unconstrained model but with less parameters. Second, when the agent learns about the level of economic growth only (constrained model with θ = 0), the goodness of fit decreases, which confirms that learning about long-run risk describes more accurately the dynamics of the observed fundamentals than learning about the level of the expected growth rate. An interpretation of our finding is that learning about the unobservable degree of longrun risk is at least as (if not more) important than learning about the expected economic growth rate itself, for which quarterly forecasts are readily available. Learning about the level of economic growth is certainly valuable, but only if the growth estimates of professional forecasters are of poor quality, which, according to our estimation, is rather unlikely. Moreover, a key uncertainty that agents face is not whether the economy is in recession (or in expansion), as given by the growth forecasts, but how long this state will last. Learning about long-run risk helps agents deal exactly with this type of uncertainty. Finally, with learning about longrun risk, the estimated persistence of the expected growth can vary over time, which is not the case when agents learn about the level of expected economic growth. To better grasp the levels and the time-variation in the main variables, Table reports 14

Mean Standard deviation Minimum Maximum f.61% 1.70% 3.75% 6.45% l 1.83 10 4 1.73 10 4 1.36 10 4 5.14 10 4 ν l 4.00 10 4.14 10 8 4.00 10 4 4.00 10 4 λ 0.037 0.037 0.11 0.083 ν λ 0.04 0.006 0.00 0.040 Table : Descriptive statistics of the main variables. This table reports the descriptive statistics of the main variables in the economy. The growth rate forecast f t is the sate of the economy, l reflects the estimated bias in forecast, and ν l to its uncertainty. Then, the filter related to learning about long-run risk is given by λ, whereas ν λ captures the uncertainty about long-run risk. some descriptive statistics. The growth rate forecast f t equals.6%, on average, and ranges between 3.8 to 6.5%. In contrast, the estimated deviation in the forecast l t is close to zero over the sample period, thereby confirming the view that professional forecasts are not systematically biased. 10 This result explains why learning about the level of the expected growth rate is not relevant when such forecasts are available. In contrast, the filter of longrun risk λ t varies strongly over time, approximately between 0.1 to 0.08. As a result, the mean-reversion speed of expected consumption growth, which is equal to θ λ t + λ, varies approximately between 0.66 to 0.86. Finally, the uncertainty about this parameter, ν λ,t, ranges between 0.0 and 0.04. Overall, these results suggest that long-run risk clearly fluctuates over time and across economic conditions. This finding thus stands in contrast to the existing asset pricing literature, which typically considers economies with constant long-run risk. Regarding the preference parameters, we fix the risk aversion to γ = 10, the elasticity of intertemporal substitution (EIS) to ψ =, and the subjective discount factor to β = 0.03. This choice of parameters is consistent with the asset pricing literature (e.g., Ai, Croce, Diercks, and Li, 015). In what follows, we use this calibration to describe the asset pricing implications of our model with learning. 4 Numerical Illustration This section provides a numerical analysis based on our calibration. Our structural estimation indicates that learning exclusively about long-run risk offers the best model fit to macroeconomic data. Based on this finding, we hereafter focus our analysis on the case θ = 1, which implies that the model only depends on the state variables f t, λ t, and ν λ,t. We first discuss the behavior of the price-dividend ratio in subsection 4.1. Then, subsec- 10 Because the estimated values of l t and 1 θ are close to 0, the uncertainty associated with the filter ν l becomes almost constant over time (see Equation 1). 15

4 ν λ = ν λ,0 ν λ = ν λ,0 ν λ = 0 4 ν λ = ν λ,0 ν λ = ν λ,0 ν λ = 0 3.9 3.9 3.8 3.8 0 0.0 0.04 0.06 Consumption growth f t 0.1 0.05 0 0.05 0.1 Estimated λ t Figure 1: Behavior of the log price-dividend ratio. For the left graph, we fix λ t = 0. For the right graph, we fix f t = f. Unless otherwise specified, the calibration used is provided in Table 1 and Section 3. tions 4.-4.3 analyze the implications of learning about long-run risk on the volatility of asset returns, the risk premium in the economy, the equilibrium risk-free rate, and the risk-return trade-off. To clearly highlight the properties of a model with learning about long-run risk, we compare our predictions with those coming from a model without uncertainty about long-run risk, i.e., when ν λ is set to zero. 11 4.1 Log price-dividend ratio The starting point of our analysis is the behavior of the log price-dividend ratio with respect to the main state variables, which we illustrate in Figure 1. The left panel shows that the log of the price-dividend ratio I increases, almost linearly, with economic conditions, as captured by the growth forecast f t. Hence, I f /I > 0. The right panel demonstrates that the log price-dividend ratio increases with the filter of the persistence parameter λ t, which implies I λ/i > 0, and thus decreases with the agent s estimate of long-run risk. Both panels suggest a decreasing relationship between the price-dividend ratio and uncertainty about long-run risk ν λ, which yields I νλ /I < 0. Hence, the price-dividend ratio increases with expected consumption growth, decreases with uncertainty about long-run risk, and decreases with the persistence of expected consumption growth. These findings, which we will also validate empirically in Section 5, numer- 11 Note that the results of the model with learning about the level of expected consumption growth only (i.e., θ = 0) are discussed in Appendix D. They are qualitatively the same as those of the traditional long-run risk model. 16

ically confirm our Conjecture 1. Eventually, Figure 1 suggests that learning about long-run risk should be an important driver of asset prices, as it exerts a strong influence on the log price-dividend ratio. 4. Stock return volatility and equity risk premium We turn now to the analysis of the volatility of asset returns and the equity risk premium. Following Equation (4) in the case θ = 1, stock market volatility is given by σ t = σ δ + ( σ f I f I + ( f f t )ν λ,t σ f while the equity risk premium, as determined by Equation (5), becomes ( µ t r t = γσδ I f + (1 φ) σ f I + ( f f t )ν λ,t σ f ) I λ, (6) I ) I λ. (7) I The first component of stock return volatility captures the uncertainty about consumption growth, while the first term of the equity risk premium represents the compensation for this risk, which is also present in the CRRA case. When the agent prefers early resolution of uncertainty, we have φ < 1 and the equity risk premium now incorporates an additional part, which consists of the same quadratic term as the one driving stock return volatility in (6). This term captures, on one hand, the compensation for long-run risk in the economy (Bansal and Yaron, 004) and, on the other hand, the uncertainty regarding the degree of long-run risk. This last component is specific to our model, as it is directly associated with learning about long-run risk. We study the properties of stock return volatility and the equity risk premium in the top and bottom panels of Figure, respectively. The left panels depict the relations of these moments with the growth forecast f t, setting the filter λ t at zero. The right panels present the relations these moments with the filter of the mean-reversion speed λ t, where the growth forecast f t has been fixed at its long-run level f. All panels report values for various levels of uncertainty about long-run risk. When the growth forecast is at its long-term level (f t = f), the model demonstrates that uncertainty about long-run risk magnifies stock return volatility and increases the equity risk premium. The reason is that incomplete information about the persistence in expected growth represents a new source of risk to investors. Thus, uncertainty about long-run risk affects asset prices even though agents are not currently learning about this risk. Indeed, under such economic conditions, consumption growth shocks remain uninformative for estimating 17

Stock return volatility 0.06 0.04 0.0 ν λ = ν λ,0 ν λ = ν λ,0 ν λ = 0 0.035 ν λ = ν λ,0 ν λ = ν λ,0 ν λ = 0 0.03 0.05 0.0 Equity risk premium 0 0.06 0.04 0.0 0 0 0.0 0.04 0.06 ν λ = ν λ,0 ν λ = ν λ,0 ν λ = 0 0 0.0 0.04 0.06 Consumption growth f t 0.0 0.015 0.01 0.005 0.1 0.05 0 0.05 0.1 0.1 0.05 0 0.05 0.1 Estimated λ t ν λ = ν λ,0 ν λ = ν λ,0 ν λ = 0 Figure : Stock return volatility and equity risk premium with learning about long-run risk. For the left plots, we fix λ t = 0. For the right plots, we fix f t = f. The dashed vertical line in the left plots corresponds to f t = f. Unless otherwise specified, the calibration used is provided in Table 1 and Section 3. the mean-reverting speed in expected consumption growth (see Equation 10), which implies that the second term in brackets in (6) and (7) becomes null. Departing now from the special case f t = f, we find that stock return volatility and the equity risk premium vary across economic conditions in presence of uncertainty about long-run risk. The left panels of Figure show that stock return volatility and the equity risk premium become negatively related with the growth forecast. In comparison, without uncertainty about long-run risk (ν λ = 0), these moments remain constant, as the second term in brackets in (6) and (7) vanishes. 1 This is the result of the long-run risk model (without 1 The only indirect channel through which equity volatility and risk premium could potentially depend on 18

stochastic volatility), which features complete information about the persistence parameter (Bansal and Yaron, 004). Hence, learning about long-run risk induces key asset pricing moments to be time-varying and counter-cyclical. Our main theoretical prediction is that the sensitivity of stock return volatility and equity risk premium to economic conditions crucially depends on the degree of uncertainty about long-run risk. Positive uncertainty activates the second term in brackets in (6) and (7), which in turn alters the response of stock returns to expected consumption growth shocks. The intuition behind this mechanism is as follows. In bad economic times (f t < f), the agent perceives positive (negative) shocks as very good (bad) news. Focusing on positive shocks, not only these shocks signal higher consumption growth in the future, but the agent now also expects less long-run risk and a shorter recession. Similarly, negative shocks during bad times imply lower future consumption growth, stronger long-run risk, and a recession that will last longer than previously expected. Overall, learning about long-run risk amplifies the sensitivity of asset returns to economic shocks during bad times, which increases stock return volatility and the equity risk premium. By contrast, learning about long-run risk has the opposite impact in good economic times. The presence of uncertainty about long-run risk reduces stock return volatility and equity risk premium during periods of high expected consumption growth (f t > f), because positive (negative) expected growth shocks become moderately good (bad) news only. To see that, positive shocks always signal higher consumption growth in the future and a longer expansion but, at the same time, induce the agent to believe that there is more long-run risk. Similarly, negative expected growth shocks signal lower consumption growth in the future and a shorter expansion, but also reduces the degree of long-run risk. Agent s updating of the long-run risk parameter thus partially offsets the effect of the initial shock when f t > f, reducing stock return volatility and the equity risk premium in good times. 13 We confirm this intuition by showing that stock return volatility and the equity risk premium strongly and negatively depend on the estimated persistence parameter λ t, as displayed in the right panels of Figure. Stock return volatility and the equity risk premium are thus higher when expected consumption growth is more persistent, which is in line with Bansal and Yaron (004). Furthermore, Figure highlights an interaction between the level and the f t is through the term I f /I. With our calibration, this effect is negligible. In fact, this effect is zero with a standard log-linear approximation (Bansal and Yaron, 004). 13 It is reasonable to expect that positive shocks during good times also signal a longer economic boom than previously expected, whereas negative shocks also signal a shorter economic boom. These expectations partially offset the effect of learning about long-run risk in good times. However, we find that the magnitude of this short-term effect is negligible when compared with the effect of changes in the persistence of consumption growth. Consequently, investors always dislike an increase in the persistence of consumption growth, because it is associated with greater long-run risk, and ultimately the term I λ/i remains positive at all times. Figure 4 in Appendix E clarifies this point. 19

uncertainty of the persistence parameter driving long-run risk. Specifically, the positive relation between stock return volatility (and the equity risk premium) and the perceived level of long-run risk strengthens when uncertainty about long-run risk is higher. Therefore, learning about long-run risk enriches and amplifies, rather than attenuates, the results of the traditional long-run risk model. Overall, our model generates a set of new predictions. First, uncertainty about long-run risk increases stock market volatility and the equity risk premium, on average. Second, this uncertainty induces stock market volatility and the equity risk premium to fluctuate over time and, in particular, to increase in bad times and to decrease in good times. Third, this uncertainty amplifies the sensitivity of asset pricing moments to the level of long-run risk in the economy. Consequently, incomplete information about the mean-reverting speed in consumption growth, which dictates the severity of long-run risk, has fundamental, new implications for the dynamics of asset prices. We will evaluate these predictions empirically in Section 5. 4.3 Risk-free rate and the risk-return tradeoff When agents learn about long-run risk only (i.e. θ = 1), the risk-free rate and the market price of risk satisfy r t = β + 1 ψ f t γ(1 + 1/ψ) σδ 1 ( (1 φ) I f σ f I + ( f f t )ν λ,t σ f ) I λ (8) I and m t = [ ( I γσ δ (1 φ) σ f f I + ( f f t)ν λ,t σ f I λi )]. (9) The upper panels of Figure 3 depict the behavior of the equilibrium risk-free rate in an economy with incomplete information about long-run risk. The risk-free rate increases with growth forecast f t (left panel) and decreases with long-run risk, which is negatively related to the filter λ t (right panel). Furthermore, uncertainty about long-run risk drives the risk-free rate in two ways. First, it decreases the level of the risk-free rate, on average, as risky assets become more uncertain. Second, it amplifies procyclicality in the risk-free rate, particularly in bad times, as the risk-free rate becomes more sensitive to changes in economic conditions. This effect can be seen through the last term of (8), which is magnified when f t < f. Hence, learning about long-run risk helps explain the low levels of risk-free rate observed during economic recessions. The lower panels of Figure 3 illustrate the model predictions regarding the Sharpe ratio. 0